619m2b72b7/evaluation/rankeval/lm1-619m-2b7-results_lm-eval_global_step5111_2023-01-23-19-58-52_1shots.json

{
    "results": {
        "anli_r1": {
            "acc": 0.324,
            "acc_stderr": 0.01480686473373886
        },
        "anli_r2": {
            "acc": 0.321,
            "acc_stderr": 0.014770821817934649
        },
        "anli_r3": {
            "acc": 0.3308333333333333,
            "acc_stderr": 0.013588208070709
        },
        "cb": {
            "acc": 0.48214285714285715,
            "acc_stderr": 0.06737697508644648,
            "f1": 0.3421052631578947
        },
        "copa": {
            "acc": 0.58,
            "acc_stderr": 0.04960449637488583
        },
        "hellaswag": {
            "acc": 0.26687910774746065,
            "acc_stderr": 0.004414246720076111,
            "acc_norm": 0.27394941246763593,
            "acc_norm_stderr": 0.00445071867355267
        },
        "rte": {
            "acc": 0.5090252707581228,
            "acc_stderr": 0.030091559826331334
        },
        "winogrande": {
            "acc": 0.4972375690607735,
            "acc_stderr": 0.014052271211616441
        },
        "storycloze_2016": {
            "acc": 0.5360769641902726,
            "acc_stderr": 0.01153229486915312
        },
        "boolq": {
            "acc": 0.5333333333333333,
            "acc_stderr": 0.008725599880049204
        },
        "arc_easy": {
            "acc": 0.37163299663299665,
            "acc_stderr": 0.009915897123658778,
            "acc_norm": 0.3442760942760943,
            "acc_norm_stderr": 0.009749495321590817
        },
        "arc_challenge": {
            "acc": 0.1680887372013652,
            "acc_stderr": 0.01092771504612486,
            "acc_norm": 0.19539249146757678,
            "acc_norm_stderr": 0.01158690718995291
        },
        "sciq": {
            "acc": 0.658,
            "acc_stderr": 0.01500870618212173,
            "acc_norm": 0.611,
            "acc_norm_stderr": 0.015424555647308493
        },
        "piqa": {
            "acc": 0.5903155603917302,
            "acc_stderr": 0.011473932007187606,
            "acc_norm": 0.5837867247007617,
            "acc_norm_stderr": 0.011500864675166568
        }
    },
    "versions": {
        "anli_r1": 0,
        "anli_r2": 0,
        "anli_r3": 0,
        "cb": 1,
        "copa": 0,
        "hellaswag": 0,
        "rte": 0,
        "winogrande": 0,
        "storycloze_2016": 0,
        "boolq": 1,
        "arc_easy": 0,
        "arc_challenge": 0,
        "sciq": 0,
        "piqa": 0
    }
}